mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-12-11 06:55:27 +00:00
Code Issues Packages Projects Releases Wiki Activity

Bug fix regarding bias

Browse Source
This commit is contained in:
Daniel Povey 2022-12-22 17:35:29 +08:00
parent a0b2276f68
commit 1dbe1e4086
1 changed files with 6 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
egs/librispeech/ASR/pruned_transducer_stateless7/scaling.py
View File

@ -432,8 +432,8 @@ class MaxEigLimiterFunction(torch.autograd.Function):
class BasicNormFunction(torch.autograd.Function):
# This computes:
# scales = torch.mean((x + bias) ** 2, keepdim=True) + eps.exp()
# return x * scales
# scales = torch.mean((x - bias) ** 2, keepdim=True) + eps.exp()
# return (x - bias) * scales
# (after unsqueezing the bias), but it does it in a memory-efficient way so that
# it can just store the returned value (chances are, this will also be needed for
# some other reason, related to the next operation, so we can save memory).
@ -448,8 +448,8 @@ class BasicNormFunction(torch.autograd.Function):
ctx.channel_dim = channel_dim
for _ in range(channel_dim + 1, x.ndim):
bias = bias.unsqueeze(-1)
scales = (torch.mean((x + bias) ** 2, dim=channel_dim, keepdim=True) + eps.exp()) ** -0.5
ans = x * scales - bias
scales = (torch.mean((x - bias) ** 2, dim=channel_dim, keepdim=True) + eps.exp()) ** -0.5
ans = x * scales
ctx.save_for_backward(ans if store_output_for_backprop else x,
scales, bias, eps)
return ans
@ -470,8 +470,8 @@ class BasicNormFunction(torch.autograd.Function):
eps.requires_grad = True
with torch.enable_grad():
# recompute scales from x, bias and eps.
scales = (torch.mean((x + bias) ** 2, dim=ctx.channel_dim, keepdim=True) + eps.exp()) ** -0.5
ans = x * scales - bias
scales = (torch.mean((x - bias) ** 2, dim=ctx.channel_dim, keepdim=True) + eps.exp()) ** -0.5
ans = x * scales
ans.backward(gradient=ans_grad)
return x.grad, bias.grad.flatten(), eps.grad, None, None